Hugging Face's logo

lealealy
/
invrobot

Robotics
JAX
orbax
Model card Files
xet
 Community

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

TinyWAM InvRobot

TinyWAM action-policy checkpoint for the InvRobot three-camera, 14-DoF setup. The repository contains only inference assets and a minimal inference runtime; no dataset or training implementation is included.

Assets

  • invrobot_step30000/: bf16 TinyWAM parameters at step 30,000.
  • wan_vae/: Wan VAE used to encode the current RGB observation.
  • dataset_stats.json: normalization constants bound to this checkpoint.
  • prompt_embeddings/put_water_flosser_in_box_close_lid.npz: checkpoint-aligned UMT5 context for the packaged example, with self-describing provenance metadata.
  • weights_manifest.json: byte sizes and SHA-256 for every required inference asset.
  • TinyWAM-InvRobot-runtime.tar.gz: inference-only JAX runtime and a real smoke sample.

The model consumes a uint8[384,320,3] RGB canvas, current float32[14] state, and a precomputed float32[128,4096] UMT5 context. It returns float32[32,14] absolute joint targets. The 22 GB UMT5 encoder is deliberately not bundled; callers supply its embedding, and both the runtime smoke sample and the standalone prompt_embeddings/ asset already contain one.

Prompt embedding provenance

The bundled prompt embedding was produced with the tokenizer/ and text_encoder/ subfolders of Wan-AI/Wan2.1-T2V-1.3B-Diffusers, pinned to revision 0fad780a534b6463e45facd96134c9f345acfa5b. The encoder is text_encoder/, loaded as transformers.UMT5EncoderModel (google/umt5-xxl architecture); the matching tokenizer/ uses max length 128 with padding, truncation, and special tokens enabled.

It encodes this raw instruction:

Put the water flosser into the box and close the lid.

through this complete prompt:

A video recorded from a robot's point of view executing the following instruction: Put the water flosser into the box and close the lid.

The NPZ contains context (float32[128,4096]), context_mask (bool[128], all true), and JSON metadata. See prompt_embeddings/README.md for a loading snippet. No UMT5 model files are duplicated in this repository.

Download and verify 

hf download lealealy/invrobot TinyWAM-InvRobot-runtime.tar.gz --local-dir .
tar -xzf TinyWAM-InvRobot-runtime.tar.gz
cd TinyWAM-InvRobot-runtime-20260817
python -m pip install -U "huggingface_hub[hf_xet]"
python scripts/download_weights.py --repo-id lealealy/invrobot
python scripts/verify_weights.py

See the runtime README.md for environment pins and the inference command.

Validation

  • External asset manifest: 60 data/weight objects verified by SHA-256.
  • Runtime syntax and import smoke: passed.
  • Orbax checkpoint and VAE restore: passed.
  • Real CUDA inference: passed on an RTX 5090 with JAX/JAXLIB 0.11.0.
  • Smoke output: shape [32,14], all values finite.
  • Existing InvRobot open-loop Gate 1 record: passed at step 30,000.

This validation is open-loop only. It is not evidence of closed-loop robot performance or safety.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
Robotics
loading